Hector Ortega
Senior Vice President
AI is already part of public sector work, whether entities planned for it or not.
People are using tools like ChatGPT, Copilot, Gemini, and other generative AI platforms to draft emails, summarize meetings, analyze documents, and speed up routine tasks.
The problem is that a lot of this is happening without formal approval, training, or oversight.
That’s shadow AI.
Shadow AI is when employees use AI tools for work without the organization knowing about it or having a say in how those tools are used. It might start with something simple, like asking a chatbot to clean up a memo or summarize a report. But if that prompt includes sensitive information — and in public sector work, it often does — the risk can climb quickly.
For public entities, shadow AI is one of the biggest risks people tend to miss when they talk about AI. It’s not always the flashy use case that causes trouble. Sometimes it’s the everyday shortcut.
Why shadow AI matters
Public entities handle a lot of sensitive information: citizen records, student files, police reports, personnel data, financial information, health-related information, and more. If someone enters that information into an unapproved AI tool, the agency may lose control over where the data goes, how it is stored, and whether it is used to train the model.
That can create real privacy, security, and compliance problems.
Unlike an AI tool that has gone through review and approval, shadow AI is often invisible until something goes wrong. There may be no procurement review, no legal review, no security review, and no record of what tool was used or what data was shared. That makes the issue harder to manage and harder to defend.
The risk is often unintentional
One reason shadow AI is so risky is that employees usually are not trying to create a problem. They are trying to work faster.
A staff member might upload a spreadsheet to spot trends. A manager might ask an AI tool to rewrite a performance note. A procurement or legal team member might use a chatbot to draft talking points or compare contract language.
In each case, the person using the tool may think they’re simply being productive. But if the tool is not approved for that use, or if sensitive data is shared without safeguards, the agency may face privacy and security exposure without realizing it happened.
That is what makes shadow AI so hard to manage. It is not just a technology issue. It is also an operational, legal, and cultural one.
What can go wrong
Shadow AI can create several kinds of problems:
Privacy and confidentiality risks
Sensitive information entered into an external tool may be stored, retained, or used in ways the agency did not intend. That can lead to privacy concerns, breach notification obligations, and reputational harm.
Accuracy risks
AI tools can produce incorrect or misleading outputs. If someone relies on that output for a public-facing response or an internal decision, the agency may end up sharing bad information.
Compliance risk
Public-sector work often comes with requirements around records retention, accessibility, public disclosure, and data handling. Shadow AI can make compliance harder because the tool may not fit neatly into existing workflows.
Security risk
Unapproved tools can open the door to more ways for a cyber issue to get in. They may be used from personal devices, connected to unmanaged accounts, or accessed outside approved IT environments.
Records and disclosure risk
If AI is used to draft or process materials, there may be records implications. Entities may need to think carefully about what is retained, what is disclosable, and how to reconstruct decision-making later.
The issue is not that AI is inherently risky. The issue is unmanaged use.
Why public entities are especially vulnerable
Shadow AI is a problem in every sector, but public entities face a few extra challenges.
First, many entities are under pressure to do more with less. That creates a natural incentive for people to use whatever tools help them move faster.
Second, public-sector teams often work across multiple functions and systems, which can make it harder to keep policies and controls consistent.
Third, entities handle highly sensitive information that should never be casually pasted into a third-party AI tool.
And finally, public entities are held to a high standard of transparency and accountability. If a shadow AI issue becomes public, the agency may have to explain not only what happened, but why no one knew about it sooner.
The first step: make AI use visible
The best way to reduce shadow AI risk is not to assume employees will stop using AI. That is unrealistic. A better approach is to make use visible and controlled.
Entities should start by asking:
- Which AI tools are currently approved?
- Which tools are employees already using informally?
- What types of data are allowed in those tools?
- Who is responsible for monitoring use?
- What happens if someone uses an unapproved platform?
Even asking those questions can reveal gaps between policy and reality.
An inventory of AI use cases is a good place to start. That inventory should include not only formal deployments, but also the tools employees may be using on their own. If the agency does not know where AI is being used, it cannot govern it effectively.
Policies need to be easy to follow
One common mistake is writing AI policies that are too broad or too technical. If the policy is hard to understand, employees will ignore it or misunderstand it.
A better approach is to make the rules practical and easy to remember.
Policies should clearly spell out:
- Which AI tools are approved for work use
- What types of data may not be entered into any external AI tool
- When human review is required
- How to report suspected misuse
- Where employees can go with questions
The goal is not to create bureaucracy for its own sake. It’s to help people use AI safely without guessing.
Training matters here too. People are more likely to make good decisions if they understand the risks in plain English. A short, practical training session is often more useful than a long legal memo.
Procurement and security need to be part of the conversation
Shadow AI is not just an HR issue or an IT issue. It is a cross-functional issue.
Procurement teams need to know what tools are being purchased and whether they come with the right contractual protections. Security teams need to understand whether those tools meet the organization’s standards. Legal and compliance teams need to know whether the use aligns with privacy, records, and disclosure obligations.
If these teams aren’t aligned, employees will make their own choices — and that’s often how shadow AI spreads.
Insurance is a critical part of the response
If shadow AI leads to a privacy event, data leak, security incident, or claim, insurance can be an important part of the response. The right coverage may help with response costs, legal defense, regulatory issues, vendor problems, and some of the financial losses that follow an incident.
Shadow AI can also create complicated claims because the issue may touch cyber, professional liability, public officials liability, employment practices, or general liability coverage — sometimes all at once.
That is why public entities should treat insurance as a central part of their AI risk strategy, not just a backstop. As unapproved AI use spreads, entities need coverage that is reviewed, coordinated, and tested against real-world scenarios.
Insurers are also paying closer attention to how organizations control AI use. A public entity that can show it has approved tools, clear policies, and employee training is generally in a stronger position than one that cannot. In that sense, governance and insurance work together: governance helps reduce the chance of an incident, and insurance helps manage the consequences when one happens.
The bottom line
Public entities should treat shadow AI as both a governance issue and an insurance issue. If people are using AI tools outside approved channels, the organization needs to understand which policies may respond, where exclusions could apply, and whether current coverage reflects how AI is actually being used.
The strongest approach brings employee habits, data safeguards, and insurance planning together before an incident occurs — not after.
To explore more insights on AI in the public sector, download our full report, Addressing AI Governance Gaps and Emerging Risks in the Public Sector, and learn how Marsh Agency can help you assess and manage this evolving risk.
